Desarrollo de un modelo de árboles de decisión para evaluar y optimizar el cumplimiento de estándares de aseguramiento de la calidad en el Instituto Superior Universitario Carlos Cisneros.

The objective of the present study was to design and validate an analytical system for diagnosing compliance with the 43 indicators of the 2024 model at the “Instituto Superior Universitario Carlos Cisneros.” To this end, a pre-existing DataFrame—into which qualitative assessments had already been converted under the fields “peso ideal” and “peso real”—served as the point of departure. From this dataset, a binary label, cumple (compliance), was generated based on a threshold ratio of ≥ 0.60. Using an 80/20 train–test split via train_test_split, a global decision tree (maximum depth = 3) was trained. Visualization of this tree revealed operational thresholds for “PESOS DE LA IES” (0.009 and 0.012) and “B PESOS SIN TP” (0.012 and 0.017), which were expressed as highly interpretable conditional rules. Five-fold stratified cross-validation (StratifiedKFold) yielded a mean accuracy of 97% and a recall of 100%, with only isolated false positives occurring in one fold. qualitative assessments had already been converted into the fields “peso ideal” and “peso real” served as the point of departure. From this dataset, a binary label “cumple” (ratio ≥ 0.60) was generated. Using an 80 %/20 % train–test split via train_test_split, a global decision tree (maximum depth = 3) was trained. Visualization of this tree revealed operational thresholds for PESOS DE LA IES (0.009 and 0.012) and B PESOS SIN TP (0.012 and 0.017), expressed as highly interpretable conditional rules. Five-fold stratified cross-validation (StratifiedKFold) produced a mean accuracy of 97 % and a recall of 100 %, with only isolated false positives in one fold. To preserve the identity of each indicator, one-hot encoding was applied, and a more granular decision tree (maximum depth = 5) was constructed. This model achieved 100% accuracy on the sample and uncovered exceptions among teaching-related indicators, such as “Formación académica en curso y capacitación.” With PESOS DE LA IES (37%) and B PESOS SIN TP (10%) emerging as the most influential features—followed by practical training and teaching compensation—a Random Forest comprising 50 trees (maximum depth = 5) was employed to quantify feature importances. Finally, an árbol de criteria (criteria tree) confirmed an overall improvement of +12.7 percentage points when comparing the 2023 and 2024 evaluations, while also highlighting a regression in the Profesores indicator. These results support a traffic-light signalling system and a validation based, ongoing action plan that prioritizes critical indicators in order to enhance institutional quality.

• 第一著者: Llanga Cruz, Carolina Valeria (author)
• フォーマット: masterThesis
• 言語: spa
• 出版事項: 2025
• 主題: aprendizaje supervisado; aseguramiento de la calidad; árbol de decisión; Random Forest; evaluación institucional
• オンライン･アクセス: http://dspace.unach.edu.ec/handle/51000/15770